stabilityai / stable-audio-3-small-sfx

Overview

Stable Audio 3 Small SFX is a text-to-audio model by Stability AI. It uses latent diffusion to denoise audio in a compressed latent space. A semantic-acoustic autoencoder then decodes those latents into a stereo waveform. It’s built for sound effects, so you can design foley, hits, and ambiences from text.

What you can build

• Foley packs for games, film, and animation (footsteps, cloth, impacts)
• UI and product sounds (clicks, toggles, alerts, success tones)
• Ambient beds (room tone, wind, rain, traffic, crowd wash)
• Cinematic transitions (risers, whooshes, drops, booms)
• Rapid prototypes for storyboards, animatics, and gameplay mockups

Inputs

• A written description of the sound you want, as plain English text.
• A target clip length in seconds. This model supports 1 to 120 seconds.
• An optional “prompt expansion” switch. It can rewrite short prompts into longer ones.
• A sampling-steps control. More steps can add detail, but changes taper off.
• A guidance strength control. Higher values can push closer to the prompt.

Outputs

The model returns one generated audio clip as a downloadable file. The response also includes the final prompt used and a numeric seed for that output. Under the hood, Stable Audio 3 generates 44.1 kHz stereo audio. Some deployments deliver the result as a compressed file, such as MP3.

Recommended settings

• Use 3 to 15 seconds for most one-shot SFX like hits, clicks, and whooshes.
• Use 15 to 60 seconds for ambiences and looping beds.
• Keep sampling steps low for iteration. The distilled checkpoints look strong at 8 steps.
• Turn prompt expansion on when your prompt is short or missing key details.
• Start guidance strength near 1. Raise it only when prompt adherence is weak.

Limitations

• It’s not designed for speech or voice generation.
• Prompts in languages other than English can perform worse.
• Maximum clip length is 120 seconds per generation.
• Highly detailed prompts can still produce artifacts like warble or smeared transients.
• Very dynamic mixes may need post work (trim, fade, normalize, and EQ).
• Vague prompts can produce generic textures instead of clear, named events.
• Low-quality or noisy source audio can hurt results when you use editing workflows.

Safety & compliance

• Follow the Stability AI Community License terms for this model.
• The pipeline uses a T5Gemma text component, which has its own use restrictions.
• Don’t generate content that violates laws, platform rules, or rights you don’t own.
• Treat outputs as generated media. Review them before publishing or shipping.
• Some deployments include an audio safety checker. It may block certain prompts.